[Qemu-devel] [RFC] mmap of BAR0 fails for ivshmem device

[Cam Macdonell]

Hello,

A bug was reported to me regarding mmaping of BAR0 in ivshmem.  Indeed the mmap fails.  This bug will affect those using the ivshmem server as BAR0 contains the registers for sending and receiving interrupts.  It does not affect those mapping just the shared memory region.

I have bisected to a patch from 3.12

commit 7314e613d5ff9f0934f7a0f74ed7973b903315d1

Author: Linus Torvalds <address@hidden>

Date:   Tue Oct 29 10:21:34 2013 -0700

    Fix a few incorrectly checked [io_]remap_pfn_range() calls

    Nico Golde reports a few straggling uses of [io_]remap_pfn_range() that

    really should use the vm_iomap_memory() helper.  This trivially converts

    two of them to the helper, and comments about why the third one really

    needs to continue to use remap_pfn_range(), and adds the missing size

    check.

    Reported-by: Nico Golde <address@hidden>

    Cc: address@hidden

    Signed-off-by: Linus Torvalds <address@hidden.

diff --git a/drivers/uio/uio.c b/drivers/uio/uio.c

index ba47563..0e808cf 100644

--- a/drivers/uio/uio.c

+++ b/drivers/uio/uio.c

@@ -642,16 +642,29 @@ static int uio_mmap_physical(struct vm_area_struct *vma)

 {

        struct uio_device *idev = vma->vm_private_data;

        int mi = uio_find_mem_index(vma);

+       struct uio_mem *mem;

        if (mi < 0)

                return -EINVAL;

+       mem = idev->info->mem + mi;

-       vma->vm_ops = &uio_physical_vm_ops;

+       if (vma->vm_end - vma->vm_start > mem->size)

+               return -EINVAL;

<snip>

The last two lines shown above that check the length of the vm area cause the mmap to fail because ivshmem's BAR0 is only 256 bytes.

One possible fix is to increase the size of BAR0 to the size of a page.  Of course, I'd prefer to be able to fix this from my uio driver, but I'm not sure that is possible given the patch above changes the generic uio code.  Advice is welcome.

Finally, I apologize for not catching this bug earlier.  It's an effect of not having the uio driver in the kernel.  To avoid this in future, I will work to get the UIO ivshmem driver into the kernel.

Sincerely,
Cam